Google's PreviewDiff: critic-guided search over diffusion latents beats Best-of-N on image and video
google · hf · 2026-10-01
Google introduced PreviewDiff, a training-free test-time search method that turns diffusion sampling into multimodal critic-guided search over intermediate latents.
- At chosen denoising checkpoints, it decodes partial previews, has a multimodal judge score and critique them, then branches over semantic prompt edits and locally re-noised continuations — spending verifier compute while the sample is still editable.
- Motivation: diffusion models struggle with compositional details (object counts, attribute binding, spatial relations), and Best-of-N can't repair a failing trajectory.
- Across image and video benchmarks, PreviewDiff consistently beats budget-matched Best-of-N and strong scalar-search baselines; earlier interventions and wider search yield the largest gains.
More from Multimodal
- Runway Research Q&A: Real-Time Models, Generative Interfaces and What's Next for Robotics — runwayml · 2026-10-01
- NVIDIA's LongLive-Plug distills video generation into plug-and-play LoRAs for 54 models — _akhaliq · 2026-10-01
- User has Opus 5.5 write a music video responding to p(doom) vids — LilijoySkySeeker · 2026-10-01
- Free Tesseract extension brings design and motion graphics into ChatGPT/Claude — aziz4ai · 2026-10-01
- ComfyUI SeedVR2 upscaler v1.5.7 tames DiT VRAM spikes, runs TensorRT under 6GB — simmol · 2026-10-01
- Commentary: Big tech dissolved music's context, making 'Personal AI Music' possible — pixlpa · 2026-10-01